dsh-code-review

DSH Plugin · One-click code review with Claude Code as the reviewer agent: the header "评审" button runs the claude CLI headless in your workspace, with the model route pointed at a cheap model (AGICTO's flash) — so you get Claude Code's agentic loop at flash-model cost.

中文文档

A persistent web plugin for DeepSeek Harness (DSH) that automates the "develop in DSH, review with Claude Code" workflow: click the 评审 button in the session header, and the host runs Claude Code (claude -p) in the session's workspace with ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN / ANTHROPIC_MODEL pointing at your provider (e.g. AGICTO) and a cheap model (e.g. deepseek-v4-flash).

The core value is Claude Code the agent's capability — its own tool loop, planning, and judgment drive the review; the underlying model is whatever you configure (flash for cost). The severity-graded report is written back to the workspace (review-report-<timestamp>.md) and shown in the UI.

How it works

[评审 button] --sessionId--> Host /review/run
      │
      ├─ subprocess.spawn("claude -p <review prompt>", {
      │     cwd: <session workspace>,
      │     env: {
      │       ANTHROPIC_BASE_URL:  https://api.agicto.cn/v1,   // AGICTO Anthropic-compatible API
      │       ANTHROPIC_AUTH_TOKEN: <AGICTO key>,
      │       ANTHROPIC_MODEL:      deepseek-v4-flash,          // flash model
      │       CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: "1" })
      │
      └─ Claude Code agent: runs git diff, reads context, writes severity-graded report
              │
              ├─ report → workspace/review-report-<ts>.md
              └─ report → UI (client polls /review/status)

• Agent = Claude Code (claude CLI, e.g. installed via npm i -g @anthropic-ai/claude-code); its model route is redirected via ANTHROPIC_* env vars.
• Model = your choice — AGICTO serves Claude models too, but the point is you can use a cheap flash model underneath while keeping Claude Code's agent behavior.
• Review is read-only: the prompt forbids modifying files or running state-changing commands.
• Reviews run in the background (up to 5 min); the client polls status until done.

Install

dsh plugin --profile web add https://github.com/Badegg404/dsh-code-review.git
# restart dsh web afterwards

Or manual: copy to ~/.dsh/profiles/web/plugins/dsh-code-review/, add "@Badegg404/dsh-code-review": "file:plugins/dsh-code-review" to the profile's dependencies and the package name to dsh.profile.bundles, then pnpm install + restart.

Prerequisite: claude (Claude Code) on PATH or at the configured path (npm i -g @anthropic-ai/claude-code).

Usage

1. Open a session (the review button lives in the session header, next to the title).
2. Click 评审 → optionally type a per-run instruction in 本次要求 (e.g. "focus on security"); under 配置 you can edit the 角色设定（系统提示词） (default role) and the agent/provider/model fields → 保存配置.
3. Click 发起评审. The host runs Claude Code headless in the workspace; it reviews the uncommitted changes.
4. The report appears in the panel when done, and a copy is written to review-report-<timestamp>.md in the workspace.

Configuration

Stored in the DSH credentials store (~/.dsh/.credentials.yaml, mode 0600):

Ref	Default	Meaning
REVIEW_PROVIDER	agicto	Provider route; also selects the default base URL
REVIEW_MODEL	(empty)	Underlying model Claude Code calls, e.g. deepseek-v4-flash
REVIEW_CLAUDE_BIN	~/.npm-global/bin/claude	Path to the claude CLI
REVIEW_BASE_URL	provider default	Anthropic-compatible API base (https://api.agicto.cn/v1)
REVIEW_API_KEY	AGICTO_API_KEY	API key; falls back to the AGICTO credential
REVIEW_AGENT	claude	Local agent to invoke: claude (Claude Code) or custom (arbitrary command)
REVIEW_CUSTOM_CMD	(empty)	Custom-agent shell command; $REVIEW_PROMPT / $REVIEW_CWD env available
REVIEW_SYSTEM_PROMPT	built-in	Role / system prompt (角色设定); {cwd} placeholder

Routes

All loopback-only (/review/*):

Route	Purpose
GET /review/status?session=<id>	config + job state / report
POST /review/config	save provider / model / claude path / base URL / key
GET /review/models?provider=<p>	advertised models (datalist)
POST /review/run	start a local-agent review (userPrompt optional)

License

MIT